SPSS uses kmeans, two-stage clustering and RFM model to study the behavior law data of borrowers and lenders in P2P network finance

Full text link ：http://tecdat.cn/?p=27831 

The source of the original text is ： The official account of the tribal public

 

With P2P The transaction volume of online financial platforms surged , Its transaction data cannot be fully and effectively utilized . Introduce cluster analysis into P2P In the management of online financial platform , Cluster analysis technology is used to analyze P2P Analyze the existing data of online financial platform , And then for the borrower 、 The service provided by the payer and the manager becomes P2P New issues faced by the network financial platform in the development process .

In view of the above problems and needs , This paper hopes to apply clustering analysis technology to P2P Online financial platform to explore the behavior law of borrowers and lenders , Thus, it is beneficial to the construction and development of the platform .

Clustering analysis

Definition of cluster analysis

Clustering analysis , It is an important research content in knowledge discovery , It is also called group analysis class , Simply put, it is a collection of elements with similar characteristics . clustering , It is to gather the elements with high similarity , Final , Form several subsets .

Algorithm and process of cluster analysis      

Clustering algorithm is the main embodiment of the superiority of clustering technology , Scalability of the algorithm 、 Ability to handle different attributes 、 Clustering ability for arbitrary shapes 、 Processing ability of noise data 、 Insensitive to the order of input records 、 High dimensional 、 Constraint based clustering, interpretability and usability can measure the quality of the algorithm .

Division method ： The partition method is to divide a given large amount of data into multiple groups or clusters according to certain rules or different partition methods , among , Each group or cluster generally contains at least one set of data , Different types of data can only belong to different groups , There are obvious differences between each group .

Hierarchical approach ： Hierarchical method for clustering analysis is to divide the data into several groups to form a tree structure , It can also be divided into top-down splitting algorithm and bottom-up condensing algorithm according to different ways of constructing numbers .

Density based approach ： This method refers to the method of continuous clustering through the density of adjacent local areas exceeding a certain threshold , in other words , Each given region will contain a certain number of points , Thus, some outliers can be filtered out through this method , Improve the efficiency of data analysis .

Typical clustering analysis process generally includes data ( Or called samples or patterns ) Get ready 、 Feature selection and feature extraction 、 Proximity calculation 、 clustering ( Or grouping )、 Evaluate the effectiveness of clustering results .

Cluster analysis process ：

1. Data preparation process ： That is, data preprocessing , Including feature standardization and dimension reduction .
2. Feature selection process ： Select the most effective feature from the initial feature and store it in the vector .
3. Feature extraction process ： Through the transformation of the selected features, a new prominent feature is formed .
4. clustering （ Or grouping ）： First, select a certain distance function of the appropriate feature type （ Or construct a new distance function ） Measure the closeness ; Then perform clustering or grouping .
5. Evaluation of clustering results ： It refers to the evaluation of clustering results . If the result is satisfactory , It's over ; If you are not satisfied, you need to adjust the above “ feature extraction ” link , Until satisfied .

Cluster analysis of borrower behavior     

Research data description

The data in this paper comes from the historical transaction information in the platform background database , Including loan related information and lender information .

The status of platform loans is divided into ： The audit failed 、 Stream label （ It refers to that the bidding period of a loan list has expired , But the loan was not fully raised , That is, the loan failed ）、 Successful borrowing （ It means that the loan is full , The loan relationship has been established ）.

After data screening , The samples of the final study include 999 Loan list . among ,248 There are articles that fail to pass the audit ;209 The bar is the flow mark ;542 Successful loan ,169 The loan has been repaid . The total amount of successful transactions reached 3090.93 Ten thousand yuan .

 

chart ： Sample data composition

According to the overall research data , The proportion of loans that failed to pass the examination reached 24.8%, The proportion of flow indicators reaches 20.9%, Only 54.3%, It can be seen that the loan success rate needs to be improved .

The purpose of cluster analysis

because P2P The online lending platform has a low threshold 、 Features such as less restrictions , This tends to exacerbate P2P The risk of online lending , therefore , Its credit system construction is very important . Its credit system is to determine the borrower's credit score according to the borrower's authentication information and the loan repayment situation , According to credit points, it is divided into AA、A、B、C、D、E、HR Seven credit ratings , among AA Class represents the highest credit rating , It means that the borrower is active in lending on the platform and has good credit , Low credit risk ; Then gradually decrease ,HR Class indicates the lowest credit rating 、 Borrowers with high credit risk .

Model design of cluster analysis   

Borrower's overall data statistics

This paper extracts 923 There are no duplicate and valid borrower information .923 Among the borrowers AA Level only 1 position , and A Level also only 1 position , and B Class and C There are relatively many classes , Most of all D、E、HR Class accounts for 98%.

 

As can be seen from the table , Among the borrowers AA、A、B、C、D、E、HR7 The number of successful loans per capita of class is greater than the number of standard loans per capita . Generally speaking, the average value , The lower the borrower's credit rating , The average number of flow marks will gradually increase , The average number of successful loans will gradually decrease . From this we can see that , The higher the credit rating of the borrower , The higher the probability of successful borrowing . But in E Class borrowers , In case of special circumstances , The average number of successful borrowings of its borrowers is greater than D Class borrowers , The average number of stream marks is less than D Class borrowers , thus , It can be seen that the classification is unreasonable . On the other hand , In terms of cardinality , Due to the high credit rating AA、A、B、C The base of class I borrowers is small , Although the loan success rate is very high , But the total number of borrowings is far less than E、HR Class borrowers . This explains , The higher the credit rating, the more successful the loan will be , And the lower the credit rating, the less successful the loan will be . The higher the credit rating with Ding Jie , The number of successes and failures of borrowers will be higher, and the conclusion is similar .

Data preparation and clustering variable selection

Through the integration of borrower information in the data source , Get user activity data , Including the user's borrowing times 、 Number of successful borrowings 、 Credit rating 、 Total loan amount and other information . In order to understand the different behavior patterns and characteristics of platform users , Here, the borrower is selected ID Number 、 Average borrowing limit 、 Total number of borrowings 、 Loan success times 、 Credit rating 、 Lend points （ The lending points obtained by the borrower as the lender ） As a clustering variable .

Application and implementation of clustering analysis

This article USES the SPSS Yes 923 Cluster analysis on the data of effective borrowers . among , Select credit rating as the classification variable , Because the sample data has both continuous variables and classified variables , therefore , This paper uses two-stage clustering . And SPSS Provided in KMeans The difference between clustering method and hierarchical clustering analysis is , The two-stage clustering method uses log maximum likelihood estimation to measure the distance between classes , And according to Schwarz Bayes criterion (BIC) or Akaike Information rules (AIC) And other indicators automatically determine the optimal number of clusters .

utilize SPSS The clustering process is as follows ：

1. Import data files into SPSS in .

Click on “ confirm ” Button , Get the results , Here's the picture .

chart The result of second-order cluster analysis

adopt SPSS Two stage clustering method , Borrowers are divided into 4 class , give the result as follows ：

The composition and characteristics of each category of users are shown in the table ：

It can be concluded from the table , The two-stage cluster analysis does not completely classify borrowers according to the credit rating set by the platform itself , Instead, it digs out more accurate behavior information of borrowers .

In the first category ,HR The proportion of graded borrowers is the largest ,D、E second , The total borrowing times of such borrowers are 4 Highest in class , But the average borrowing limit 、 The number of successful loans and lending points are 4 The lowest class , It can be seen that although this type of borrower is more active on the platform, its degree of trust is very low , There is a serious risk of fraud .

In the second category , D Borrowers with class a credit rating account for 100%, Although the average loan amount and the total number of loans are not the highest , However, its loan success times and loan points are the highest among the four categories , It can be seen that , Although such borrowers are not very active on the platform as borrowers , However, the proportion of full standard loans issued by it is very high , And they are often active on the platform as lenders .

Cluster analysis of lender behavior     

The purpose of cluster analysis

This paper uses the method of cluster analysis , Yes P2P In the online lending platform, lenders segment customers , So as to find out the category of lenders , In the end P2P There is a reasonable and accurate classification of lenders in the online lending platform .

Customer segmentation

be based on RFM Analysis of customer classification principle of the model

There are many methods of customer segmentation , But the final criterion to measure whether the method is suitable should be the accuracy of the segmentation results and the matching degree with the enterprise management .

RFM Model is one of the commonly used methods of customer segmentation , As a quantitative analysis model , It is generally used for prediction and analysis before implementing marketing activities . among ,R（recenty） Last consumption , It refers to the time interval between the latest consumption and the current date , Theoretically speaking , The closer the customer spent last time , The more likely you are to respond to the goods or services provided immediately ;F（frequency） Consumption frequency , It refers to a certain period of time , The number of times customers spend , Usually , The more customers spend , The more loyal you are , That means we can have more market share by increasing the number of customers' consumption ;M（monetary） Consumption amount , It refers to a certain period of time , The total amount of customer consumption , Consumption amount is the pillar of all data rate reports , You can also verify “ Pareto's law ”—— company 80% The income comes from 20% Customers , Through the consumption amount, we can see which are the key customers , Contribute the most to the turnover of the company .

according to P2P Characteristics of online lending platform , take RFM The indicators should be changed accordingly , As shown in the table .

be based on K-Means Customer classification of clustering algorithm

K-Means Clustering method , Also known as K- Mean clustering method is widely used in clustering algorithm based on partition .K-Means The algorithm is based on the number of categories entered k value , All objects in cluster analysis are divided into k Groups , There is a high similarity between objects in each group . This article takes K-Means Clustering is a tool , Weighted RFM For measure , by P2P Lenders are classified in the online lending platform , The basic idea is ：

1） take RFM The third index is standardized , Before weighting, the data needs to be standardized . use Ri、Fi、Mi Respectively refers to the standardized lender i Of R、F、M value .

also ,Ri=(RM-R)/(RM- RN),Fi=(F-FN)/(FM-FN),Mi=(M-MN)/(MM-MN) （1）

among ,RM、 RN Among the lenders R Maximum and minimum values of ,FM、FN Among the lenders F Maximum and minimum values of ,MM、MN Sub table for lenders M Maximum and minimum values of .

2） Determine the number of clustering groups k.

3） Weight the indicators , And make use of K-Means Clustering method to get k Class lenders .

4） For each type of lender RFM Mean and total RFM Compare the mean , Finally, determine the customer type of each type of lender .

The application of clustering analysis in the customer segmentation of lenders

This paper extracts 500 Borrower information that is not repeated and valid . Relevant important data are shown in the table below .

Use K-means When clustering the mean , You need to judge the number of clusters in advance .RFM In the model , The classification of lenders' customers is through each customer category RFM Average and total RFM Compare the average value to determine , And the comparison of a single indicator can only have 2 In this case ： Greater than ( be equal to ) Or less than the average , So there may be 2×2×2=8 Species category , therefore , In this paper, the number of clusters is determined as 8 individual . First , According to the formula （1） For lenders R、F、M Values are normalized , Then use SPSS19.0 Software for standardized lenders R、F、M Value for K-Means Clustering analysis .

Finally get 8 Type of lender , And will 8 Lender like R、F、M Mean and total R、F、M Mean comparison , among “↑” Is greater than the average ,“↓” Is less than the average , The results are as follows .

Discussion on the user training strategy of the platform

Now , At home P2P Online lending platforms are developing rapidly , But most of them ignore the part of user training , Especially for lenders . The platform has not formed a perfect and effective mechanism and strategy in user training , In a short time , It may not reflect the importance of user training , But it is likely to cause the loss of high-quality users for a long time .

Through the cluster analysis of front-end borrowers, we can see , There are some key development customers among the borrowers , It may not have a high credit rating, but it borrows a lot on the platform and can repay on time , Such borrowers can be regarded as key development targets , Provide them with some encouragement and preferential policies . For high-quality borrowers that already exist on the platform , We can launch corresponding loan projects according to their actual needs .

---

The most popular insights

1.R Language k-Shape Algorithm stock price time series clustering

2.R Language clusters cities hierarchically based on temperature 、kmean clustering 、 Principal component analysis and Voronoi chart

3.R Language is used to analyze the time series data of power load K-medoids Clustering modeling and GAM Return to

4.r Language iris iris Hierarchical clustering of data sets

5.Python Monte Carlo K-Means Cluster practice

6. use R Mining and clustering website comment text

7.R Language KMEANS Mean clustering and hierarchical clustering ： Visualization of similarities and differences in quality of life happiness in Asian countries and regions

8.PYTHON User churn data mining ： Establish logistic regression 、XGBOOST、 Random forests 、 Decision tree 、 Support vector machine 、 Naive Bayesian model and KMEANS Clustering user profiles

9.R Language based Keras Deep learning image classification based on small data sets